Learning Temporal Rules from State Sequences

In this paper we consider the problem of learning rules about temporal relationships between labeled time intervals. We learn these rules from a single series of such labeled intervals, which might be obtained from (multivariate) time series by extracting various features of interest, for instance segments of increasing and decreasing local trends. We seek for the identification of frequent local patterns in this state series . A temporal pattern is defined as a set of states together with their interval relationships described in terms of Allen’s interval logic, for instance “A before B, A overlaps C, C overlaps B” or equivalently “state A ends before state B starts, the gap is covered by state C”. In the spirit of association rule mining we propose an algorithm to discover frequent temporal patterns and to generate temporal rules. As an application we consider the problem of deriving local weather forecasting rules that allow us to conclude from the qualitative behaviour of the air-pressure curve to the windstrength.